Mistral-7B-Instruct-v0.2-ORPO-SALT

This model is a fine-tuned version of mistralai/Mistral-7B-Instruct-v0.2 on the dpo_mix_en and the bct_non_cot_dpo_1000 datasets. It achieves the following results on the evaluation set:

• Loss: 0.8630
• Rewards/chosen: -0.0798
• Rewards/rejected: -0.1033
• Rewards/accuracies: 0.5618
• Rewards/margins: 0.0235
• Logps/rejected: -1.0328
• Logps/chosen: -0.7975
• Logits/rejected: -2.8558
• Logits/chosen: -2.8473
• Sft Loss: 0.7975
• Odds Ratio Loss: 0.6547

Model description

More information needed

Intended uses & limitations

More information needed

Training and evaluation data

More information needed

Training procedure

Training hyperparameters

The following hyperparameters were used during training:

• learning_rate: 5e-06
• train_batch_size: 2
• eval_batch_size: 2
• seed: 42
• gradient_accumulation_steps: 8
• total_train_batch_size: 16
• optimizer: Adam with betas=(0.9,0.999) and epsilon=1e-08
• lr_scheduler_type: cosine
• lr_scheduler_warmup_steps: 0.1
• num_epochs: 3.0

Training results

Training Loss	Epoch	Step	Validation Loss	Rewards/chosen	Rewards/rejected	Rewards/accuracies	Rewards/margins	Logps/rejected	Logps/chosen	Logits/rejected	Logits/chosen	Sft Loss	Odds Ratio Loss
0.93	0.8082	500	0.8928	-0.0828	-0.1057	0.5773	0.0229	-1.0567	-0.8278	-2.8442	-2.8368	0.8278	0.6496
0.7571	1.6165	1000	0.8679	-0.0803	-0.1034	0.5700	0.0231	-1.0337	-0.8028	-2.8729	-2.8644	0.8028	0.6510
0.8035	2.4247	1500	0.8630	-0.0798	-0.1033	0.5618	0.0235	-1.0328	-0.7975	-2.8558	-2.8473	0.7975	0.6547

Framework versions

• PEFT 0.10.0
• Transformers 4.40.1
• Pytorch 2.3.0
• Datasets 2.19.0
• Tokenizers 0.19.1